What a generative AI assistant actually does to support productivity — the field data

Most numbers circulating about AI and customer support productivity come from the vendor selling the tool. One of the few large field studies not funded by a chatbot company comes from economists Erik Brynjolfsson, Danielle Li and Lindsey Raymond, in Generative AI at Work, a National Bureau of Economic Research working paper later published in the Quarterly Journal of Economics. It tracked 5,179 customer support agents at a software company as a generative AI conversational assistant was rolled out gradually across teams, which let the researchers compare agents with and without access at the same point in time.

The average result, and the result that actually matters

The headline number is a 14% average increase in issues resolved per hour. The number worth remembering is what sits underneath it: a 34% productivity gain for novice and lower-skilled agents, against almost no measurable effect for experienced, highly skilled ones. The researchers’ interpretation is that the assistant effectively distributes the conversational patterns and phrasing of the team’s best performers to everyone else, moving newer agents down the experience curve faster rather than making already-strong agents faster still.

What that means for a support team of any size

The practical implication is that a blanket rollout measured only in aggregate hides where the value actually lands. If your best agents already handle the tricky cases well, an assistant will not obviously speed them up — but if half your team is newer or still ramping up, that is where the resolution-per-hour number moves. The study also found improved customer sentiment and higher agent retention alongside the productivity gain — a detail easy to miss when the pitch is only about speed.

• →Segment the measurement by seniority before rolling out — an average hides the fact that most of the gain concentrates among newer or lower-skilled agents.
• →Treat the assistant as a way to spread what your best agents already know, not a replacement for training material.
• →Track customer sentiment and retention alongside resolution speed — the study found both moved, not just the throughput number vendors usually lead with.